Abstract
Intelligent transportation systems (ITS) supported by Internet of Things (IoT) devices and unmanned aerial vehicles (UAVs) are vulnerable to delay-related cyber anomalies, including wireless interference, jamming, and buffer overflows, which degrade service quality and threaten operational safety. A delay-aware intrusion detection system (IDS) is developed using Federated Learning (FL) with local Reinforcement Learning (RL) adaptation to provide adaptive and privacy-preserving threat detection. Local RL agents learn anomaly classification policies from sequential traffic features, while a federated aggregation mechanism combines model updates without exposing raw data. The framework utilizes deep temporal models to capture both short-term and long-term delay dynamics, along with an optimized reward function that balances detection accuracy, false alarms, latency, and communication cost. Evaluation in a three-site IoT-UAV simulation shows that the proposed system improves overall detection accuracy to 95.7%, exceeds conventional centralized and federated IDS by up to 3.3 percentage points, reduces the median detection latency from 0.7 s to 0.4 s, and lowers communication overhead compared to centralized training. These results demonstrate a scalable, low-latency, and privacy-preserving security architecture capable of providing a secure solution for next-generation ITS.